AI Helps Crack 18-Year Quantum Problem as Startup Eyes Science Business

Physical Superintelligence says artificial intelligence helped its team solve the Aaronson-Ambainis conjecture, an 18-year-old quantum-computing problem, and the startup is now trying to turn that kind of breakthrough work into a business. The result, reported by TechCrunch, could narrow one path by which quantum computers might speed up certain hard problems, while also signaling how quickly AI is moving from general-purpose assistants into scientific research.

What Physical Superintelligence says it achieved

According to TechCrunch, the company used early access to Anthropic’s Fable 5.2 to help solve the Aaronson-Ambainis conjecture, a problem that had stood open for 18 years. The conjecture sits in quantum query complexity, an area that examines the limits of how much faster quantum machines can be than classical computers for specific tasks.

That matters because the result does not claim to settle quantum computing as a whole. Instead, TechCrunch reported that the proof narrows one route for quantum computers to accelerate some hard problems, but not all. In other words, the finding is more surgical than sweeping, but still notable in a field where progress often comes in narrow theoretical increments.

Who worked on it?

One of the people named in TechCrunch’s reporting is David Turturean, a recent MIT graduate who worked on the proof as technical staff at Physical Superintelligence. Turturean had deferred plans for MIT’s computer science and AI master’s program to join the company, a decision that reflects how aggressively some young researchers are choosing startup work over the traditional graduate-school path.

Scott Aaronson, whose name is attached to the conjecture, told TechCrunch that he does not fully understand the proof, but believes it is correct. He also said the problem had been among the last major open questions in quantum query complexity, underscoring why the announcement drew attention from both theorists and the growing community around AI-assisted science.

Why it matters

The immediate significance is mathematical, but the broader significance is commercial. Physical Superintelligence is not presenting the result as a one-off stunt; it is framing it as evidence that AI can become part of a repeatable research workflow. That puts the startup in the middle of a larger race to productize scientific discovery tools rather than simply build chatbots or coding assistants.

TechCrunch reported that current revenue at Physical Superintelligence comes from data-center optimization work, where the company models heat, power, water and sound for data-center siting and performance. That is a practical business line with clear customers. The longer-term pitch is more ambitious: use the same research infrastructure to convert theoretical physics breakthroughs into technologies.

For readers, investors and researchers, the key question is no longer whether AI can help with narrow scientific tasks. It is whether those tools can be assembled into a business that repeatedly produces verifiable research output, and whether the market will pay for that capability before the standards around proof, disclosure and peer review are fully settled.

How did AI become part of the proof process?

The answer, based on TechCrunch’s reporting, is a mix of model access, human expertise and careful validation. Physical Superintelligence said the team used Anthropic’s Fable 5.2 during the work, then formalized the proof in Lean, a software language used to mechanically verify mathematical arguments. The company also explained the result to a team of researchers it shared the proof with, adding another layer of scrutiny before making the work public.

That workflow is important because AI-generated or AI-assisted mathematical claims can sound impressive long before they are trustworthy. Formal proof tools help close that gap by forcing the argument into a machine-checkable form. In this case, the company’s use of Lean suggests it was not relying on a loose language model answer, but on a chain of verification that included formalization and expert review.

Still, the reporting leaves room for caution. TechCrunch noted that Scott Aaronson does not fully understand the proof himself, even as he believes it is correct. That tension is familiar in frontier science: a result can appear convincing to experts while still inviting independent verification from the broader field.

Background: the 18-year problem and the new AI research push

The Aaronson-Ambainis conjecture dates to 2008, which means it remained open for 18 years before Physical Superintelligence’s claim. The conjecture is part of the theory landscape that tries to pin down where quantum computers should, and should not, have an advantage over classical machines. Because the problem is about the limits of speedups, its solution is useful less as a gadget breakthrough than as a boundary marker for the field.

That makes it a good test case for the current wave of AI research systems. AI labs and startups are increasingly building specialized systems for science rather than only general-purpose assistants, and the physical sciences are an especially demanding target. As the fact sheet notes, physics is harder to automate than math because proposed results must match experiments in the real world. A correct proof is not the same thing as a working physical device.

Physical Superintelligence appears to understand that distinction. CEO Matt Pines told TechCrunch that the company builds research systems with harnesses, tools, memory and simulation infrastructure. He also said the company’s longer-term goal is to turn theoretical-physics breakthroughs into technologies. In his telling, the startup is not selling a chatbot; it is building a stack for scientific work.

What PSI says it is building

Pines described an effective research system as something layered and operational, not merely conversational. In the company’s view, the useful system needs tools for running tasks, memory for continuity, harnesses for controlled execution and simulation infrastructure for testing ideas. That architecture fits the demands of science far better than a generic prompt box.

“An effective kind of research system should have lots of things related to harnesses, lots related to tools, memory systems, simulation infrastructure”

That comment, reported by TechCrunch, is revealing because it frames AI as part of a broader research machine. It also hints at how such products may be sold: not as a single model, but as an integrated environment that can help scientists narrow hypotheses, check arguments and move faster toward testable claims.

By the numbers

The timeline around the result helps show why the announcement drew attention across both AI and quantum-computing circles.

What happened after the proof was ready?

TechCrunch reported that Physical Superintelligence held the result until Anthropic’s model was near release, then shared it with relevant experts. That timing suggests a careful, staged disclosure strategy rather than a rush to generate attention. The company seems to have wanted both a credible technical package and the right moment to present it publicly.

The involvement of outside experts matters because scientific claims rise or fall on scrutiny. Jim Halverson of Northeastern University, who advises Physical Superintelligence, told TechCrunch that the company chose transparency and expert contact as interim best practices. He also said, “Best practices are still emerging,” a line that captures the current state of AI-assisted science better than any corporate slogan could.

“Best practices are still emerging”

That is the crux of the story. The proof may be real, but the field still lacks settled norms for how AI-generated or AI-assisted discoveries should be validated, credited and published. For startups, that leaves room for innovation. For everyone else, it leaves room for uncertainty.

Why this result is different from a typical AI headline

This is not a story about a model writing prose, passing a benchmark or answering questions faster than before. It is about AI helping a startup push through a long-standing theoretical barrier in quantum computing, with the result then being translated into a business narrative. That combination makes the story more consequential than a typical product launch.

It also raises a subtler issue: when AI helps produce a proof, where does the credit sit? The fact sheet makes clear that human researchers were still central. Turturean worked on the proof, experts reviewed it, Lean formalized it and Aaronson weighed in on its plausibility. AI may have accelerated the process, but it did not replace the need for human judgment.

That balance could shape how other labs present future work. If Physical Superintelligence can show that AI-assisted research leads to reproducible gains, more startups may package similar systems for chemistry, materials science, physics or applied engineering. If it cannot, the result may be remembered as an impressive exception rather than a durable operating model.

How are commercial and scientific goals intersecting?

They are intersecting directly. Physical Superintelligence’s current revenue comes from data-center optimization, not from selling the conjecture result itself. But the company’s stated ambition is to build a platform for theoretical physics breakthroughs and, ultimately, technologies that emerge from them. That is a classic frontier-startup play: use a practical service to fund a more speculative research engine.

Discovered Materials appears to be moving in a related direction. According to TechCrunch, it launched a $1 million AI science challenge focused on superconductivity and heat dissipation. The appearance of a cash prize in a related scientific domain shows that the market is starting to recognize AI-driven discovery as a contest worth funding, not just a research curiosity.

For the industry, this matters because it could pull more capital toward specialized scientific systems. Not every startup will need to solve an 18-year quantum problem to attract attention. But the bar is rising for teams claiming they can use AI to improve research. They will increasingly need evidence, not just vision.

What happens next

The next phase is verification. TechCrunch’s report makes clear that one open question is whether the proof will withstand full independent scrutiny from the wider community. Even with Lean formalization and expert review, the final test is whether other mathematicians and quantum-computing researchers accept the argument under standard checks.

Another question is downstream science. Because the conjecture concerns limits on quantum speedups, it may influence how researchers think about which problems truly benefit from quantum methods. But the fact sheet also notes that any physics implications would still need experiments, since models cannot override real-world constraints.

There is also a business question. Physical Superintelligence has said it wants to commercialize scientific-discovery systems, but the precise product shape remains unclear. Will the company keep leaning on data-center optimization while building a broader research platform, or will it spin out more focused tools for physicists and other scientists?

One more signal to watch is the reaction from peers. Next week, an MIT summit is set to gather physicists to discuss the new pace of AI-driven science. That meeting could become a useful barometer for whether the field sees this as an isolated success, a model for future work or a warning that standards need to catch up quickly.

What the reaction says about the field

Scott Aaronson’s reaction, as reported by TechCrunch, captures the emotional range such a result can provoke. He said he felt satisfied, awestruck, disappointed and fatigued by the news. That mix makes sense: a long-open problem is finally resolved, but by a team using methods and tools that complicate the usual idea of mathematical discovery.

He also told TechCrunch, “I was surprised, honestly, to see it missing from OpenAI’s Mathocalypse dump a few weeks ago.” The remark is notable not because it changes the result, but because it shows how fast the competitive map around AI research is shifting. Important work can now surface from unexpected places and at unexpected moments.

For Physical Superintelligence, that is an opportunity and a challenge. The opportunity is obvious: if the company can keep producing credible results, it may become a reference point for AI-assisted science. The challenge is that every future claim will likely face sharper scrutiny than the last.

Bottom line

Physical Superintelligence says AI helped it solve an 18-year quantum-computing conjecture, and it is using that result to argue for a new kind of research business. The accomplishment, reported by TechCrunch, sits at the intersection of formal mathematics, AI tooling and startup ambition. Its scientific value will depend on independent verification. Its commercial value will depend on whether the company can turn a breakthrough into a repeatable product.

For now, the story is less about a single proof than about a changing research economy. AI is no longer just assisting with text or code. It is entering the workflows where science is made, and companies like Physical Superintelligence are racing to define what that means before the field defines it for them.